We present fully apodized and perfectly vertical surface grating couplers in 300 nm silicon waveguides. We achieve ultrahigh coupling efficiency of -0.35 dB at 1550 nm for an optical fiber with 10.4 µm mode-fiber-diameter. To this end, we followed a two stage process in which, we first optimized a pool of periodic grating couplers using machine learning techniques, and then apodized them over 100 parameters using the gradient based adjoint- method. Using a simple fabrication tolerance analysis, we also show that segment variations mostly causes a wavelength shift for the maximum coupling efficiency of the apodized grating couplers, similar to those typically observed for periodic grating couplers.
Modern design of photonic devices is quickly and steadily departing from classical geometries to focus more and more on non-trivial structures and metamaterials. These devices are governed by a multitude of parameters and the optimal design requires to simultaneously consider different figure of merits. In this invited talk we will present our recent work on the application of machine learning tools to the multi-objective optimization of multi-parameter photonic devices. In particular, we will demonstrate the potentiality of dimensionality reduction for the analysis of the complex design space of subwavelength metamaterials devices.
Enabled by technological improvements, photonic devices and circuits are becoming increasingly more complex. Non-trivial geometries are designed to reduce device footprint, improve performance, and introduce novel functionalities. However, the number of design variables required to properly represent these geometries quickly grows, limiting the effectiveness of classical design approaches. Moreover, parameters are often strongly interdependent, restricting the use of sequential optimizations or independent parameter sweeps. Although several optimization techniques can be effective for multi-parameter design, they commonly allow to optimize for a single or a handful designs and the optimization process needs to be repeated if new performance criteria are introduced. In contrast to classical design approaches, the in uences of the design parameters remain hidden as well as the general behavior of the design space. In this paper we present an extension of our recent work on the application of machine learning pattern recognition to the design of multi-parameter photonic devices. In particular, we propose using a combination of local optimization based on the adjoint method and the use of dimensionality reduction. Adjoint optimization is used multiple times to generate a small set of different designs with high performance. Dimensionality reduction is applied to analyze the relationship between these degenerate designs and identify a lower-dimensional design sub-space that includes all alternative good designs. This sub-space can be mapped for any performance criteria thus enabling informed decisions based on the relative priorities of all relevant performance specifications. As a proof of concept, we demonstrate a ten-parameter design of an integrated photonic power splitter using silicon-on-insulator technology. We identify a region of possible high performance design solutions and select two design candidates either maximizing the splitter efficiency or minimizing back-reflection.
Miniaturized silicon photonics spectrometers have great potential for mass market applications like medicine and hazard detection. However, the performance of state-of-the-art silicon spectrometers is limited by fabrication imperfections and temperature variations.
In this work, we present a fundamentally new strategy that combines machine learning algorithms and on-chip spatial heterodyne Fourier-transform spectroscopy to identify specific absorption features operated under a wide range of temperatures in the presence of fabrication imperfections. We experimentally show differentiation of four different input spectra with unknown temperature variations as large as 10 °C. This is about 100x increase in operational range, compared to state-of-the-art retrieval techniques.
The performance and functionality of integrated photonic devices can be enhanced by using complex structures controlled by a large number of design variables. However, the optimization of such high-dimensional structures is challenging, often limiting their realization. Global optimization algorithms and artificial neural networks are increasingly used to tackle these problems. Although these are exciting new developments, the outcome is a single optimized design meeting particular performance objectives selected upfront. The influences of the various design parameters remain hidden. Here we report on our strategy of using machine learning pattern recognition techniques to create a methodology for building the global performance map of a high-dimensional design space. As an example and demonstration, we study the design of a vertical grating coupler consisting of silicon and subwavelength metamaterial segments. We show how the relationship between designs with comparable primary performance can be clearly revealed by identifying the minimum number of characterizing parameters that defines the subspace of good designs, significantly scaling down the complexity of the problem. Moreover, the subspace can be identified using only a small number of good design solutions. We reveal design areas with comparable fiber coupling efficiency but with significant differences in other performance criteria, such as back-reflections, tolerance to fabrication uncertainty and minimum feature size. This novel approach provides the designer with a global perspective of the design space, enabling informed decisions based on the relative priorities of all relevant performance specifications and figures-of-merits for a particular application. Insights from the mapping exercise also inspired new design structures with enhanced characteristics.
Integrated nanophotonic component design processes are often constrained by computational resources. Advances in simulation and optimization tools have allowed more efficient exploration of larger design spaces. These developments reduce the time-consuming and intuition-limited effort of encoding physical insights into the design structure. However, we argue that efficient optimization is only part of the solution to tackle larger multi-parameter design spaces. Finding patterns in such a space can be more valuable than identifying the individual optima alone. This is particularly true when transitioning from simulation to real device fabrication, where considerations such as tolerance to fabrication imperfections, bandwidth, etc. take an important role but are ignored at the optimization stage. The elucidation of patterns in a complex design space enables efficient identification of designs addressing these additional considerations. As an example, in this presentation we demonstrate how limited data collected from the optimization process of a multisegment vertical grating coupler can be used to identify such patterns through the application of machine learning techniques. The identified patterns, some more interpretable than others, can be used in multiple ways: from speeding up the remaining optimization process itself to gaining insight into the properties of an interesting subset of designs. Together those insights offer a significantly clearer picture of the design space and form the basis for making much more informed decisions on the final designs to be fabricated.
Machine-assisted design of integrated photonic devices (e.g. through optimization and inverse design methods) is opening the possibility of exploring very large design spaces, novel functionalities and non-intuitive geometries. These methods are generally used to optimize performance figures-of-merit. On the other hand, the effect of manufacturing variability remains a fundamental challenge since small fabrication errors can have a significant impact on light propagation, especially in high-index-contrast platforms. Brute-force analysis of these variabilities during the main optimization process can become prohibitive, since a large number of simulations would be required. To this purpose, efficient stochastic techniques integrated in the design cycle allow to quickly assess the performance robustness and the expected fabrication yield of each tentative device generated by the optimization. In this invited talk we present an overview of the recent advances in the implementation of stochastic techniques in photonics, focusing in particular on stochastic spectral methods that have been regarded as a promising alternative to the classical Monte Carlo method. Polynomial chaos expansion techniques generate so called surrogate models by means of an orthogonal set of polynomials to efficiently represent the dependence of a function to statistical variabilities. They achieve a considerable reduction of the simulation time compared to Monte Carlo, at least for mid-scale problems, making feasible the incorporation of tolerance analysis and yield optimization within the photonic design flow.